Korean exchange rate forecasts using Bayesian variable selection

ABSTRACT

Using Bayesian variable selection, we demonstrate that economic variables forecast Korea-US exchange rates better than random walk or random walk with drift model at a short horizon. It implies that the failure of out-of-sample exchange rate forecasts is due to the uncertainties associated with selecting proper predictors, rather than the lack of relationship between the exchange rate and its theoretical determinants. Our results also suggest that time-variant and asymmetric weights on predictors should be taken into account to understand exchange rates dynamics. (JEL classification: C11, C53, F31)

KEYWORDS:

• Exchange rates forecasting
• out-of-sample predictability
• Bayesian MCMC algorithm
• parameter heterogeneity

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1. Wieland and Wolters (2013) provide a detailed review on how forecasts are used in policymaking in US and Europe.

2. Cheung and Chinn (2001) and Fratzscher et al. (2015) provide empirical evidence of this argument by detecting the time-variant weights on the exchange rate predictors.

3. Rossi (2013) state that ‘a consensus in the literature is that the Taylor-rule and net foreign assets fundamentals have more out-of-sample predictive content that traditional fundamentals.’ However, she also points that the performance of the exchange rate forecasts depends on the choice of predictor, forecast horizon, sample period, and forecast evaluation method.

4. In this paper, we use 15 variables which are conventionally used in the literature. Compared to the Wright (2008)’s BMA (eight variables for monthly data, nine variables for quarterly data) or Byrne, Korobilis, and Ribeiro (2018)’s DMA and DMS (six variables), BVS enable us to involve more predictors in the predictive regression than other methods do.

5. Korobilis (2013) extends the stochastic variable selection method to generic linear and nonlinear models and provides forecasts of the unemployment rate, the inflation rate, and interest rates.

6. Mizrach (1992), Diebold and Nason (1990), and Engel et al. (2015) also use the lagged exchange rate changes as a predictor in their setting.

7. See Section 2 for the references therein.

8. Lee (2007) states that the Korean government announced a plan to liberalize all foreign exchange transaction in June 1998.

9. Rossi (2013) states that the short window enables parameters to adapt more quickly to possible structure changes. The estimated parameter in the smaller sample, however, is less efficient.

10. Kilian (1999) has argued that there is no evidence of higher predictability at longer horizons.

11. Figure 1 is based on the results when the driftless random walk is the benchmark. Using a random walk with drift as the benchmark, we gain very similar plots. It is available upon request.

12. Here, we report the estimated based on the direct forecast method. We also estimate the predictors’ probability using iterated multi-step forecasts and gain consistent results. Plots for iterated multi-step forecasts are available upon request.

13. Lee (2018) shows that US economic policy uncertainty plays more important role in Korea foreign exchange markets than Korean economic policy uncertainty does.

14. It is well known that longer-horizon predictive analyses are prone to inference bias from using overlapping data. However, our aim of this paper is finding out the best performance model of out-of-sample forecasts and no statistical inference is involved. We also concern that non-overlapping data might lose information and the small sample size is not enough to analyze by BVS for 17 years. Thus, we report the results using overlapping data.

15. For the model parameters, we evaluate the inefficient factor and Geweke-p values by implementing the convergence of the MCMC. For an inefficient factor, we choose the maximum autocorrelation order to be 200. As a result, the highest inefficient factor is 62.169 and the lowest Geweke-p is 0.6456 among all model parameters. We can conclude that our MCMC algorithm sufficiently converged.

16. We do not present the corresponding results. These results are available upon request.

17. There is an alternative way to control data-scaling: semi-automatic approach proposed by George, Sun, and Ni (2008). The performance of out-of-sample forecasts using the semi-automatic approach does not differ from the result of hierarchical prior.

18. We perform both direct and multi-step forecasts when the prior v₀ and δ₀ are equal to 0.01. This is a fairly non-informative inverse-gamma prior to fully reflect data information. We observe that the predictive power is very similar to our previous empirical results. For the sake of brevity, we do not present the corresponding results.